enable_input_require_grads is called twice with use_gradient_checkpointing=True and bnb quantisation; is there any effect?

[nzw0301]

As in the quantization tutorial, we need to call the following lines to use peft with quantization.

from peft import prepare_model_for_kbit_training
model = prepare_model_for_kbit_training(model)

In the method with use_gradient_checkpointing=True, which is a default value,
model.enable_input_require_grads is called in

peft/src/peft/utils/other.py

Lines 111 to 122 in 8452d71

	if (loaded_in_kbit or is_gptq_quantized or is_aqlm_quantized) and use_gradient_checkpointing:
	# When having `use_reentrant=False` + gradient_checkpointing, there is no need for this hack
	if "use_reentrant" not in gradient_checkpointing_kwargs or gradient_checkpointing_kwargs["use_reentrant"]:
	# For backward compatibility
	if hasattr(model, "enable_input_require_grads"):
	model.enable_input_require_grads()
	else:
	def make_inputs_require_grad(module, input, output):
	output.requires_grad_(True)
	model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)

.

Then almost same logic is executed in _prepare_model_for_gradient_checkpointing via get_peft_model. So in my understanding, model.enable_input_require_grads can be called twice.

[BenjaminBossan]

Yes, calling this twice should not change the result.

[nzw0301]

Thank you for your response!

[divisionblur]

What is the function of the line of code model.enable_input_require_grads ?

[divisionblur]

Is it also necessary to set this during the pre-training phase?

[BenjaminBossan]

The explanation can be found here:

https://github.com/huggingface/transformers/blob/0f67ba1d741d65b07d549daf4ee157609ce4f9c1/src/transformers/modeling_utils.py#L2273-L2278

For pre-training, you would typically not use PEFT (instead updating all parameters) so it should not be required.